Comet Browser by Perplexity vs Sup AI

“The direct competitors here are Arc Browser's AI features, Dia from The Browser Company, Google's built-in Gemini integration in Chrome, and frankly just using Perplexity in a tab. The scenario where Comet breaks is the moment a user hits a site with aggressive bot detection, a multi-step OAuth flow, or a form that requires human verification — and that's the majority of 'book an appointment' use cases in the real world. My prediction for what kills this in 12 months: Google ships Gemini-native task execution in Chrome and the 3.5 billion people who already have Chrome installed don't download a new browser for a feature they get for free. For Comet to earn a ship, it needs to demonstrate autonomous task completion on a real-world benchmark — not a curated demo set — and show completion rates above 70% on genuinely complex multi-step workflows.”

“Extraordinary claims require extraordinary evidence. A 7.41 point jump on HLE via ensembling — without publishing methodology — smells like benchmark gaming. The latency of running 339 models in parallel is also a real concern for anything other than async research tasks.”

“The thesis Comet is betting on: within three years, the browser's primary job shifts from rendering documents to executing intentions, and whoever owns the execution layer owns the session data that trains the next generation of personal agents. The dependency that has to hold is that users will switch browsers — which historically requires extraordinary activation energy, but smartphone-generation users have shown less browser loyalty than desktop users, and Perplexity already has distribution through its search product. The second-order effect that matters most isn't the time saved booking appointments; it's that Comet positions Perplexity to capture behavioral clickstream data at a scale that currently only Google holds, which becomes the actual moat. This is riding the trend of 'intent graph beats knowledge graph' and Perplexity is approximately on-time — not early enough to be alone, but not late enough to be irrelevant.”

“Model ensembling is an underexplored direction in the race to reduce hallucination. If Sup AI's approach scales, it could be more durable than fine-tuning individual models — you get the wisdom of the crowd across model families, training data, and architectures simultaneously.”

“The buyer here is the existing Perplexity Pro subscriber who is already paying $20/month and now gets a reason to make Perplexity their primary browsing context, not just a search tab — that's a defensible expansion play into a relationship they already own. The moat question is harder: browser switching costs are real but the moat isn't the browser itself, it's the behavioral data and the agent memory that accumulates over sessions, which is the right answer but requires years of retention to materialize. The stress-test that concerns me most isn't Google — it's that Perplexity's own unit economics depend on query costs, and an agentic browser that runs multi-step tasks is dramatically more expensive per session than a search query; if they can't make the margin work at scale, the Pro pricing doesn't hold.”

“The job-to-be-done as stated is 'browse the web and get things done without context-switching to an AI tool' — which is one coherent job, so the focus is there. The problem is completeness: a browser only works as a daily driver if it handles 100% of browsing tasks, and Comet launching without extension support, established sync infrastructure, password manager integration, and a mature dev tools panel means users will dual-wield Chrome and Comet for months, which is the death state for browser adoption. The product has a clear opinion — AI executes, human approves — but the onboarding question I need answered is whether a new user reaches a successful autonomous task completion in under five minutes or spends that time granting permissions and watching it fail on a CAPTCHA.”

“The HLE claim needs independent verification, but the underlying ensemble approach is architecturally sound for factual Q&A tasks. Running 339 models is expensive — pricing will be the gating factor for production use. The $10 free credit is a fair trial.”

“For creative work, ensemble outputs tend to regress toward the mean — you get the most-agreed-upon version of something, which is usually the least interesting version. This is a tool for factual accuracy, not creativity. I'd stick with a single strong model for writing.”